Nothing Special   »   [go: up one dir, main page]

WO2023206724A1 - 用于电子级二氟甲烷制备的精馏控制系统及其控制方法 - Google Patents

用于电子级二氟甲烷制备的精馏控制系统及其控制方法 Download PDF

Info

Publication number
WO2023206724A1
WO2023206724A1 PCT/CN2022/097770 CN2022097770W WO2023206724A1 WO 2023206724 A1 WO2023206724 A1 WO 2023206724A1 CN 2022097770 W CN2022097770 W CN 2022097770W WO 2023206724 A1 WO2023206724 A1 WO 2023206724A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature
feature map
distillation
predetermined time
tower
Prior art date
Application number
PCT/CN2022/097770
Other languages
English (en)
French (fr)
Inventor
张朝春
林百志
刘志强
李卫国
黄华华
邱玲
赖甜华
廖耀东
Original Assignee
福建德尔科技股份有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 福建德尔科技股份有限公司 filed Critical 福建德尔科技股份有限公司
Publication of WO2023206724A1 publication Critical patent/WO2023206724A1/zh

Links

Images

Classifications

    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B01—PHYSICAL OR CHEMICAL PROCESSES OR APPARATUS IN GENERAL
    • B01D—SEPARATION
    • B01D3/00—Distillation or related exchange processes in which liquids are contacted with gaseous media, e.g. stripping
    • B01D3/009—Distillation or related exchange processes in which liquids are contacted with gaseous media, e.g. stripping in combination with chemical reactions
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B01—PHYSICAL OR CHEMICAL PROCESSES OR APPARATUS IN GENERAL
    • B01D—SEPARATION
    • B01D3/00—Distillation or related exchange processes in which liquids are contacted with gaseous media, e.g. stripping
    • B01D3/14—Fractional distillation or use of a fractionation or rectification column
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B01—PHYSICAL OR CHEMICAL PROCESSES OR APPARATUS IN GENERAL
    • B01D—SEPARATION
    • B01D3/00—Distillation or related exchange processes in which liquids are contacted with gaseous media, e.g. stripping
    • B01D3/42—Regulation; Control
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07C—ACYCLIC OR CARBOCYCLIC COMPOUNDS
    • C07C17/00—Preparation of halogenated hydrocarbons
    • C07C17/093—Preparation of halogenated hydrocarbons by replacement by halogens
    • C07C17/20—Preparation of halogenated hydrocarbons by replacement by halogens of halogen atoms by other halogen atoms
    • C07C17/202—Preparation of halogenated hydrocarbons by replacement by halogens of halogen atoms by other halogen atoms two or more compounds being involved in the reaction
    • C07C17/206—Preparation of halogenated hydrocarbons by replacement by halogens of halogen atoms by other halogen atoms two or more compounds being involved in the reaction the other compound being HX
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07C—ACYCLIC OR CARBOCYCLIC COMPOUNDS
    • C07C17/00—Preparation of halogenated hydrocarbons
    • C07C17/38—Separation; Purification; Stabilisation; Use of additives
    • C07C17/383—Separation; Purification; Stabilisation; Use of additives by distillation

Definitions

  • the present invention relates to the field of intelligent manufacturing, and more specifically, to a distillation control system for the preparation of electronic-grade difluoromethane and a control method thereof.
  • Difluoromethane is a Freon substitute with good thermodynamic properties.
  • the ODP of R32 is 0 and the GWP value is very low.
  • the azeotropes and near-azeotropic mixtures formed by R32 and other components are regarded as the most potential substitutes for R22.
  • the liquid phase fluorination method is widely used in the production of R32.
  • the reaction mixture also contains the intermediate product R31, R23, R22, R40, R21, R143a, R50 generated by other side reactions, and entrained A small amount of raw materials R30 and HF.
  • the purity of electronic-grade difluoromethane is as high as 99.9999%, which places higher requirements on the existing difluoromethane preparation process.
  • Existing manufacturers mainly improve the purity of difluoromethane through two technical routes.
  • One technical direction is to change the preparation principle of difluoromethane (essentially a chemical direction), and the other technical direction is to optimize the purification process of difluoromethane ( Essentially a physical direction), but whether it is a chemical direction or a physical direction, it ultimately requires optimization of the purification process (especially the distillation process) to prepare electronic-grade difluoromethane that meets the purity requirements.
  • deep learning and neural networks have been widely used in computer vision, natural language processing, speech signal processing and other fields.
  • deep learning and neural networks have also shown that they are close to or even beyond human performance in areas such as image classification, object detection, semantic segmentation, and text translation.
  • Embodiments of the present application provide a rectification control system for the preparation of electronic-grade difluoromethane and a control method thereof, wherein the rectification control system for the preparation of electronic-grade difluoromethane includes a reactor and a reflux tower. , degassing tower, distillation tower and controller.
  • An artificial intelligence-based parameter control algorithm is deployed in the controller to dynamically adjust the distillation control parameters of the distillation tower based on the global situation of the distillation control system. In this way, from the perspective of optimal control To improve the purification accuracy of electronic grade difluoromethane.
  • a distillation control system for electronic-grade difluoromethane preparation which includes:
  • Reflux tower used to receive the first generated mixed gas containing difluoromethane and separate the hydrogen fluoride, the difluoromethane and the monofluorochloromethane from the generated mixed gas containing difluoromethane. To obtain the second generated mixture;
  • a degassing tower configured to receive the second product mixture and remove trifluoromethane and methane in the second product mixture to obtain a third product mixture
  • a rectification tower configured to receive the third generated mixed gas and perform rectification on the third generated mixed gas to obtain a distillation product, where the distillation product is electronic grade difluoromethane with a purity of greater than or equal to 99.9999%; as well as
  • a controller for dynamically controlling the temperature and pressure of the rectification tower based on global parameters of the rectification control system where the global parameters of the rectification control system include the pressure of the reflux tower, the reflux tower temperature, the pressure of the degassing tower, the temperature of the degassing tower, the pressure of the rectification tower and the temperature of the rectification tower.
  • the controller is used for:
  • the multiple control parameters include: the pressure of the reflux tower, the temperature of the reflux tower, the pressure of the degassing tower, the The temperature of the degassing tower, the pressure of the rectification tower and the temperature of the rectification tower;
  • the first feature vectors at each predetermined time point are two-dimensionally arranged into a feature matrix and then passed through a second convolutional neural network to obtain a second feature map;
  • the responsiveness estimate between the first feature map and the second feature map is calculated using normalization based on the characterization information relationship between the local and the global to obtain a responsive feature map, wherein the using is based on the local and global
  • the normalization of the characterization information relationship between them is to divide the logarithmic function value of the sum of the eigenvalues of each position in the first feature map and one by the sum of the eigenvalues of all positions in the second feature map. the value of the logarithmic function summed with one;
  • the response feature map is passed through a classifier to obtain a classification result.
  • the classification result is used to indicate that the pressure of the rectification tower should be increased or decreased, and the temperature of the rectification tower should be increased or decreased. Small.
  • the controller is further used to: use the first convolutional neural network using a three-dimensional convolution kernel to calculate the The gas chromatograms of the distillation products at multiple predetermined time points are encoded to obtain the first characteristic diagram;
  • H j , W j and R j represent the length, width and height of the three-dimensional convolution kernel respectively
  • m represents the number of the (i-1)th layer feature map
  • b ij is the bias
  • f represents the activation function.
  • the controller includes:
  • An embedding conversion unit configured to use the embedding layer of the encoder model containing the context of the embedding layer to respectively convert a plurality of control parameters at each of the predetermined time points into input vectors to obtain a sequence of parameter input vectors;
  • a context encoding unit configured to perform global-based context semantic encoding on the sequence of parameter input vectors using the converter of the encoder model containing the context of the embedded layer to obtain the plurality of feature vectors;
  • a cascading unit is used to cascade multiple feature vectors to obtain a first feature vector corresponding to each predetermined time point.
  • the controller is further used for:
  • the input data is respectively subjected to convolution processing, pooling processing along the channel dimension and activation processing in the forward pass of the layer to obtain the final result of the second convolutional neural network.
  • One layer generates the second feature map, wherein the input of the first layer of the second convolutional neural network is the feature matrix.
  • the controller is further used for:
  • the controller is further used to: use the classifier to process the responsiveness feature map with the following formula to generate a classification result ;
  • the formula is: softmax ⁇ (W n ,B n ):...:(W 1 ,B 1 )
  • a control method which includes:
  • the multiple control parameters include: the pressure of the reflux tower, the temperature of the reflux tower, the pressure of the degassing tower, the temperature of the degassing tower, The pressure of the distillation column and the temperature of the distillation column;
  • the first feature vectors at each predetermined time point are two-dimensionally arranged into a feature matrix and then passed through a second convolutional neural network to obtain a second feature map;
  • the responsiveness estimate between the first feature map and the second feature map is calculated using normalization based on the characterization information relationship between the local and the global to obtain a responsive feature map, wherein the using is based on the local and global
  • the normalization of the characterization information relationship between them is to divide the logarithmic function value of the sum of the eigenvalues of each position in the first feature map and one by the sum of the eigenvalues of all positions in the second feature map. the value of the logarithmic function summed with one;
  • the response feature map is passed through a classifier to obtain a classification result.
  • the classification result is used to indicate that the pressure of the rectification tower should be increased or decreased, and the temperature of the rectification tower should be increased or decreased. Small.
  • multiple control parameters at each of the predetermined time points are passed through a context encoder including an embedding layer to obtain multiple feature vectors, And concatenate multiple feature vectors to obtain the first feature vector corresponding to each predetermined time point, including:
  • Multiple feature vectors are concatenated to obtain a first feature vector corresponding to each predetermined time point.
  • the distillation control system for the preparation of electronic-grade difluoromethane includes a reactor and a reflux tower. , a degassing tower, a distillation tower and a controller, wherein an artificial intelligence-based parameter control algorithm is deployed in the controller to dynamically adjust the distillation rate of the distillation tower based on the global situation of the distillation control system. Distillation control parameters, in this way, improve the purification accuracy of electronic grade difluoromethane from the perspective of optimized control.
  • Figure 1 is a schematic block diagram of a distillation control system for the preparation of electronic-grade difluoromethane according to an embodiment of the present application.
  • FIG. 2 illustrates a block diagram of a controller in the distillation control system for electronic-grade difluoromethane preparation according to an embodiment of the present application.
  • Figure 3 is a flow chart of a control method of a distillation control system for electronic grade difluoromethane preparation according to an embodiment of the present application.
  • Figure 4 is a schematic structural diagram of a control method of a distillation control system for electronic-grade difluoromethane preparation according to an embodiment of the present application.
  • difluoromethane (R32) is a thermodynamically good alternative to Freon.
  • the ODP of R32 is 0 and the GWP value is very low.
  • the azeotropes and near-azeotropic mixtures formed by R32 and other components are regarded as the most potential substitutes for R22.
  • the liquid phase fluorination method is widely used in the production of R32.
  • the reaction mixture also contains the intermediate product R31, R23, R22, R40, R21, R143a, R50 generated by other side reactions, and entrained A small amount of raw materials R30 and HF.
  • the purity of electronic-grade difluoromethane is as high as 99.9999%, which places higher requirements on the existing difluoromethane preparation process.
  • Existing manufacturers mainly improve the purity of difluoromethane through two technical routes.
  • One technical direction is to change the preparation principle of difluoromethane (essentially a chemical direction), and the other technical direction is to optimize the purification process of difluoromethane ( Essentially a physical direction), but whether it is a chemical direction or a physical direction, it ultimately requires optimization of the purification process (especially the distillation process) to prepare electronic-grade difluoromethane that meets the purity requirements.
  • deep learning and neural networks have been widely used in computer vision, natural language processing, speech signal processing and other fields.
  • deep learning and neural networks have also shown that they are close to or even beyond human performance in areas such as image classification, object detection, semantic segmentation, and text translation.
  • multiple control parameters of the rectification system at multiple predetermined time points are first obtained.
  • the multiple control parameters include: the pressure of the reflux tower, the pressure of the reflux tower temperature, the pressure of the degassing tower, the temperature of the degassing tower, the pressure of the rectification tower and the temperature of the rectification tower, and at the same time, obtain the distillation products at the multiple predetermined time points gas chromatogram.
  • the gas chromatograms of the distillation products at the plurality of predetermined time points are passed through a first convolutional neural network using a three-dimensional convolution kernel to obtain a first feature map.
  • the first feature map using a three-dimensional convolution kernel is A convolutional neural network can effectively extract the dynamic characteristics of the distillation products.
  • multiple control parameters at each predetermined time point are passed through a context encoder including an embedding layer to obtain multiple feature vectors, and the multiple feature vectors are concatenated to obtain the first feature corresponding to each predetermined time point. vector.
  • the context encoder can perform global encoding on each parameter based on context semantics to extract high-dimensional latent features of each parameter and global high-dimensional latent features between various parameters.
  • the first feature vectors at each predetermined time point are two-dimensionally arranged into a feature matrix and then passed through the second convolutional neural network to obtain the second feature map. That is, the second convolutional neural network is used to extract each Implicit associations between parameters between time points.
  • the control result of the desired control parameters can be obtained.
  • the gas chromatography feature expressed by the first feature map F 1 is a local three-dimensional correlation of the three-dimensional convolution kernel Based on feature extraction, it focuses more on local feature expression, so it easily leads to low dependence on global responsiveness when calculating responsiveness. Based on this, it is calculated using a normalized expression based on the representation information relationship between the part and the whole, specifically:
  • the aggregation of the responsiveness of the feature locally equivalent to the feature as a whole is achieved, thereby improving the responsiveness of the feature map for the first feature map F1 pair
  • the global dependence of the expected responsiveness of the second feature map F 2 thereby improves the accuracy of the final classification.
  • this application proposes a distillation control system for the preparation of electronic-grade difluoromethane, which includes: a reactor for receiving difluoromethane and hydrogen fluoride, wherein the difluoromethane and the hydrogen fluoride are in A reaction occurs under the catalysis of a catalyst to generate a first mixed gas containing difluoromethane, and the catalyst is loaded in the reactor; a reflux tower is used to receive the first mixed gas containing difluoromethane.
  • a degassing tower for receiving the third The second product mixture is used to remove trifluoromethane and methane in the second product mixture to obtain the third product mixture; a rectification tower is used to receive the third product mixture and mix the third product mixture.
  • the gas is rectified to obtain a distillation product, which is electronic grade difluoromethane with a purity of greater than or equal to 99.9999%; and, a controller, used to: obtain multiple data of the rectification system at multiple predetermined time points.
  • control parameters include: the pressure of the reflux tower, the temperature of the reflux tower, the pressure of the degassing tower, the temperature of the degassing tower, the pressure of the rectification tower and The temperature of the distillation tower; obtaining the gas chromatograms of the distillation products at the multiple predetermined time points; passing the gas chromatograms of the distillation products at the multiple predetermined time points through using a three-dimensional convolution kernel
  • the first convolutional neural network obtains the first feature map; passing multiple control parameters at each predetermined time point through a context encoder including an embedding layer to obtain multiple feature vectors, and concatenating the multiple feature vectors to Obtain the first feature vector corresponding to each predetermined time point; arrange the first feature vectors at each predetermined time point in two dimensions into a feature matrix and then use the second convolutional neural network to obtain the second feature map; use local-based and the normalization of the characterization information relationship between the whole to calculate the responsiveness estimate between the first feature map and the second feature map to obtain the responsive
  • Figure 2 illustrates a flow chart of a distillation control system for electronic grade difluoromethane preparation according to an embodiment of the present application.
  • the distillation control system 200 for the preparation of electronic grade difluoromethane according to the embodiment of the present application includes: a reactor 210, a reflux tower 220, a degassing tower 230, a distillation tower 240 and a controller 250 .
  • the reactor 210 is used to receive dichloromethane and hydrogen fluoride, wherein the dichloromethane and the hydrogen fluoride react under the catalytic action of a catalyst to generate a first mixed gas containing difluoromethane,
  • the catalyst is packed in the reactor. That is, the reactor 210 is a place where a crude product of difluoromethane is generated through a chemical reaction.
  • Zhang Yanhong the difluoromethane is produced using a gas-phase fluorination process, and its chemical reaction process includes:
  • the difluoromethane can also be produced using other principles, which is not limited by this application.
  • the reflux tower 220 is configured to receive the first generated mixed gas containing difluoromethane and separate the hydrogen fluoride, the difluoromethane and the monofluoride from the generated mixed gas containing difluoromethane. Monochloromethane to obtain the second mixed gas. That is to say, after the first generated mixed gas containing difluoromethane output from the reactor 210 is input to the reflux tower 220, the reflux tower 220 sequentially separates HF through its acid gas separation system. and HCl.
  • the degassing tower 230 is configured to receive the second product mixture and remove trifluoromethane and methane in the second product mixture to obtain a third product mixture. Specifically, considering that trifluoromethane and methane have relatively low boiling points, after the second generated mixed gas enters the degassing tower 230, the degassing tower can use the third Different components in the secondary gas mixture have different boiling points to filter out low-boiling impurities trifluoromethane and methane (ie, R23 and R50).
  • the rectification tower 240 is configured to receive the third generated mixed gas and perform rectification on the third generated mixed gas to obtain a distillation product.
  • the distillation product is electronic grade II with a purity of greater than or equal to 99.9999%. Fluoromethane. That is to say, the distillation tower 240 is used to purify the third generated mixed gas to produce the electronic grade difluoromethane.
  • the reactor 210, the reflux tower 220, the degassing tower 230 and the rectification tower 240 can use any existing equipment to construct the rectification tower. Distillation control system. Compared with the traditional distillation control system, the inventor of the present application optimized the distillation and purification accuracy of monofluoromethane from the perspective of the control end.
  • the purity of electronic-grade difluoromethane is as high as 99.9999%, which places higher requirements on the existing difluoromethane preparation process.
  • Existing manufacturers mainly improve the purity of difluoromethane through two technical routes.
  • One technical direction is to change the preparation principle of difluoromethane (essentially a chemical direction), and the other technical direction is to optimize the purification process of difluoromethane ( Essentially a physical direction), but whether it is a chemical direction or a physical direction, it ultimately requires optimization of the purification process (especially the distillation process) to prepare electronic-grade difluoromethane that meets the purity requirements.
  • the distillation control system 200 for the preparation of electronic-grade difluoromethane further includes the controller 250, wherein a controller based on The artificial intelligence parameter control algorithm dynamically adjusts the distillation control parameters of the distillation tower based on the overall situation of the distillation control system. In this way, the efficiency of electronic grade difluoromethane is improved from the perspective of optimal control. Purified precision.
  • the controller 250 is configured to obtain multiple control parameters of the distillation system at multiple predetermined time points, where the multiple control parameters include: the parameters of the reflux tower. pressure, the temperature of the reflux tower, the pressure of the degassing tower, the temperature of the degassing tower, the pressure of the rectification tower and the temperature of the rectification tower; obtain the values of the multiple predetermined time points
  • the gas chromatogram of the distillation product passing the gas chromatogram of the distillation product at the plurality of predetermined time points through a first convolutional neural network using a three-dimensional convolution kernel to obtain a first feature map; passing each of the Multiple control parameters at a predetermined time point are passed through a context encoder including an embedding layer to obtain multiple feature vectors, and the multiple feature vectors are concatenated to obtain a first feature vector corresponding to each predetermined time point;
  • the first feature vector at the predetermined time point is two-dimensionally arranged into a feature matrix and then passed through the second convolution
  • Responsiveness estimation between a first feature map and a second feature map is used to obtain a responsive feature map, wherein the normalization based on the characterization information relationship between the local and the whole is used to calculate each feature map in the first feature map.
  • the logarithmic function value of the sum of the feature values of the positions and one is divided by the logarithmic function value of the sum of the feature values of all positions in the second feature map and the logarithmic function value of one; and, the responsiveness feature map is divided by
  • the classifier obtains a classification result, and the classification result is used to indicate that the pressure of the rectification tower should be increased or decreased, and the temperature of the rectification tower should be increased or decreased.
  • FIG. 2 illustrates a block diagram of a controller in the distillation control system for electronic-grade difluoromethane preparation according to an embodiment of the present application.
  • the controller 250 includes: a parameter acquisition unit 251, used to acquire multiple control parameters of the distillation system at multiple predetermined time points.
  • the multiple control parameters include: the reflux The pressure of the tower, the temperature of the reflux tower, the pressure of the degassing tower, the temperature of the degassing tower, the pressure of the rectification tower and the temperature of the rectification tower; the product data acquisition unit 252 uses In order to obtain the gas chromatograms of the distillation products at the plurality of predetermined time points; the convolution encoding unit 253 is used to convert the gas chromatograms of the distillation products at the plurality of predetermined time points by using a three-dimensional convolution kernel.
  • the first convolutional neural network to obtain the first feature map; the context encoding unit 254 is used to pass multiple control parameters of each predetermined time point through a context encoder including an embedding layer to obtain multiple feature vectors, and Multiple feature vectors are concatenated to obtain the first feature vector corresponding to each predetermined time point; the correlation pattern extraction unit 255 is used to two-dimensionally arrange the first feature vectors at each predetermined time point into a feature matrix and then pass The second convolutional neural network obtains the second feature map; the multi-receptive field normalization unit 256 is used to calculate the first feature map and the second feature map using normalization based on the representation information relationship between the local and the whole.
  • Responsiveness estimation between feature maps to obtain responsive feature maps wherein the normalization based on the characterization information relationship between the local and the whole is based on the feature value of each position in the first feature map and a The logarithmic function value of the sum divided by the logarithmic function value of the sum of the feature values of all positions in the second feature map and the logarithmic function value of the sum of one; and, the control result generation unit 257 is used to convert the responsive feature
  • the graph is passed through a classifier to obtain a classification result, which is used to indicate that the pressure of the rectification tower should be increased or decreased, and the temperature of the rectification tower should be increased or decreased.
  • multiple control parameters of the distillation control system and distillation product data are collected.
  • multiple control parameters of the distillation system at multiple predetermined time points can be obtained through multiple temperature and pressure sensors provided in the distillation system, and the multiple predetermined time points can be obtained through a gas chromatograph.
  • the gas chromatogram of the distillation product discharged from the distillation tower is obtained, that is, the global parameters of the distillation control system are obtained.
  • the gas chromatograms of the distillation products at the multiple predetermined time points are passed through the first step using a three-dimensional convolution kernel.
  • Convolutional neural network to obtain the first feature map. That is, in the technical solution of the present application, the gas chromatograms of the distillation products at multiple predetermined time points are further processed through the first convolutional neural network using a three-dimensional convolution kernel to extract the The local correlation features of the gas chromatogram of the distillation product in the time series dimension are used to obtain the first feature map.
  • the first convolutional neural network using a three-dimensional convolution kernel can effectively extract the dynamic change characteristics of the distillation product.
  • the process of passing the gas chromatograms of the distillation products at multiple predetermined time points through the first convolutional neural network using a three-dimensional convolution kernel to obtain the first feature map includes: : Use the following formula to pass the gas chromatograms of the distillation products at the plurality of predetermined time points through a first convolutional neural network using a three-dimensional convolution kernel to obtain a first feature map;
  • H j , W j and R j represent the length, width and height of the three-dimensional convolution kernel respectively
  • m represents the number of the (i-1)th layer feature map
  • b ij is the bias
  • f represents the activation function.
  • the context encoding unit 254 of the controller 250 is used to pass multiple control parameters of each predetermined time point through a context encoder including an embedding layer to obtain multiple features. vector, and concatenate multiple feature vectors to obtain a first feature vector corresponding to each predetermined time point.
  • a context encoder including an embedding layer to obtain multiple features. vector, and concatenate multiple feature vectors to obtain a first feature vector corresponding to each predetermined time point.
  • multiple control parameters at each predetermined time point are passed through a context encoder including an embedding layer to obtain multiple feature vectors, and the multiple feature vectors are concatenated to obtain the corresponding
  • the process of obtaining the first feature vector at each predetermined time point includes: first, using the embedding layer of the encoder model that includes the context of the embedding layer to respectively convert a plurality of control parameters at each of the predetermined time points into input vectors to Get a sequence of parameter input vectors. Then, global-based context semantic encoding is performed on the sequence of parameter input vectors using the transformer of the encoder model containing the context of the embedded layer to obtain the plurality of feature vectors. Finally, multiple feature vectors are concatenated to obtain a first feature vector corresponding to each predetermined time point.
  • the correlation pattern extraction unit 255 of the controller 250 is used to two-dimensionally arrange the first feature vectors of each predetermined time point into a feature matrix and then pass the second volume Accumulate the neural network to obtain the second feature map. That is to say, in the technical solution of the present application, after obtaining the first feature vector, the first feature vectors at each predetermined time point are two-dimensionally arranged into a feature matrix and then processed through the second convolutional neural network. Processing to extract implicit correlation features between parameters between the various time points, thereby obtaining a second feature map.
  • the input data is subjected to convolution processing, pooling processing along the channel dimension and activation processing in the forward pass of the layer through each layer of the second convolutional neural network to be processed by the The last layer of the second convolutional neural network generates the second feature map, wherein the input of the first layer of the second convolutional neural network is the feature matrix.
  • normalization based on the characterization information relationship between the local and the whole is used to calculate the first Responsiveness estimation between the feature map and the second feature map to obtain the responsive feature map, wherein the normalization based on the characterization information relationship between the local and the whole is based on each position in the first feature map
  • the logarithmic function value of the sum of the feature values and one is divided by the logarithmic function value of the sum of the feature values of all positions in the second feature map and one.
  • the first feature map and the second feature map can be further fused and passed through a classifier to obtain the desired control parameters. result.
  • the gas chromatography feature expressed by the first feature map F 1 is a local three-dimensional correlation of the three-dimensional convolution kernel Based on feature extraction, it focuses more on local feature expression, so it easily leads to low dependence on global responsiveness when calculating responsiveness. Therefore, in the technical solution of the present application, a normalized expression based on the characterization information relationship between the part and the whole is further used to calculate the responsiveness estimate between the first feature map and the second feature map to obtain the responsiveness.
  • Feature map a normalized expression based on the characterization information relationship between the part and the whole is further used to calculate the responsiveness estimate between the first feature map and the second feature map to obtain the responsiveness.
  • the responsiveness estimate between the first feature map and the second feature map is calculated using normalization based on the characterization information relationship between the part and the whole to obtain the responsiveness feature.
  • the process of mapping includes: using normalization based on the characterization information relationship between the local and the whole to calculate the responsiveness estimate between the first feature map and the second feature map using the following formula to obtain the responsive feature map;
  • control result generating unit 257 of the controller 250 is used to pass the responsive feature map through a classifier to obtain a classification result, and the classification result is used to represent the The pressure of the distillation column should be increased or decreased, and the temperature of the distillation column should be increased or decreased. That is to say, in the technical solution of the present application, after obtaining the response characteristic map, the response characteristic map is further passed through a classifier to obtain a value indicating whether the pressure of the distillation tower should be increased or decreased. Small, the classification result of the distillation tower temperature should be increased or should be reduced.
  • the classifier processes the responsiveness feature map with the following formula to generate a classification result, where the formula is: softmax ⁇ (W n ,B n ):...:( W 1 ,B 1 )
  • the distillation control system 200 for the preparation of electronic-grade difluoromethane based on the embodiment of the present application is clarified, which uses a first convolutional neural network using a three-dimensional convolution kernel to obtain all the data from multiple predetermined time points. Extract the dynamic change characteristics of the distillation product from the gas chromatogram of the distillation product, and use the context encoder to extract the high-dimensional implicit features of each control parameter at the multiple predetermined time points and the relationships between the parameters.
  • the global high-dimensional latent features further use a normalized expression based on the representation information relationship between the local and the whole to fuse the two feature information, so that by introducing the robustness around the minimization loss of the representation information to the responsiveness estimation, The aggregation of the local feature equivalent to the responsiveness of the entire feature is achieved, thereby improving the global dependence of the responsive feature map on the expected responsiveness of the first feature map to the second feature map. In turn, the accuracy of classification can be improved.
  • the controller 250 in the rectification control system 200 for the preparation of electronic grade difluoromethane according to the embodiment of the present application can be implemented in various terminal equipment, such as the rectification system for the preparation of electronic grade difluoromethane. Control algorithm server, etc.
  • the controller 250 according to the embodiment of the present application can be integrated into the terminal device as a software module and/or a hardware module.
  • the controller 250 may be a software module in the operating system of the terminal device, or may be an application program developed for the terminal device; of course, the controller 250 may also be a software module of the terminal device.
  • One of many hardware modules are examples of many hardware modules.
  • the controller 250 and the terminal device may also be separate devices, and the controller 250 may be connected to the terminal device through a wired and/or wireless network, and according to the agreed data format to transmit interactive information.
  • Figure 3 illustrates a flow chart of a control method of a distillation control system for electronic grade difluoromethane production.
  • the control method of the distillation control system for the preparation of electronic grade difluoromethane according to the embodiment of the present application includes the step of: S110, obtaining multiple controls of the distillation system at multiple predetermined time points.
  • the plurality of control parameters include: the pressure of the reflux tower, the temperature of the reflux tower, the pressure of the degassing tower, the temperature of the degassing tower, the pressure of the rectification tower and the The temperature of the distillation tower; S120, obtain the gas chromatograms of the distillation products at the multiple predetermined time points; S130, obtain the gas chromatograms of the distillation products at the multiple predetermined time points by using three-dimensional convolution The first convolutional neural network of the kernel is used to obtain the first feature map; S140, pass the multiple control parameters of each predetermined time point through the context encoder including the embedding layer to obtain multiple feature vectors, and convert the multiple feature vectors Perform cascade to obtain the first feature vector corresponding to each predetermined time point; S150, arrange the first feature vectors of each predetermined time point in two dimensions into a feature matrix and then use the second convolutional neural network to obtain the second feature vector.
  • Feature map use normalization based on the characterization information relationship between the local and the whole to calculate the responsiveness estimate between the first feature map and the second feature map to obtain the responsive feature map, wherein, the The normalization based on the characterization information relationship between the local and the whole is divided by the logarithmic function value of the sum of the feature value of each position in the first feature map and one by the logarithmic function value of all positions in the second feature map.
  • the logarithmic function value of the sum of feature values and the sum of one; and, S170 pass the responsiveness feature map through a classifier to obtain a classification result, which is used to indicate that the pressure of the distillation tower should be increased. should be increased or should be decreased, the temperature of the distillation tower should be increased or should be decreased.
  • FIG. 4 illustrates a schematic structural diagram of a control method of a distillation control system for electronic-grade difluoromethane preparation according to an embodiment of the present application.
  • the gas chromatograms of the distillation products obtained at the multiple predetermined time points are obtained (For example, P1 as shown in Figure 4)
  • a first feature map (for example, as shown in Figure 4) is obtained by using a first convolutional neural network (for example, CNN1 as shown in Figure 4) of a three-dimensional convolution kernel F1); S140, pass multiple control parameters (for example, P2 as shown in Figure 4) of each predetermined time point through a context encoder including an embedding layer (for example, as shown in Figure 4 E)
  • Obtain multiple feature vectors for example, VF1 as illustrated in Figure 4
  • concatenate the multiple feature vectors to obtain a first feature vector corresponding to each predetermined time point (for example,
  • steps S110 and S120 multiple control parameters of the rectification system at multiple predetermined time points are obtained.
  • the multiple control parameters include: the pressure of the reflux tower, the temperature of the reflux tower. , the pressure of the degassing tower, the temperature of the degassing tower, the pressure of the rectification tower and the temperature of the rectification tower, and obtain the gas phase of the rectification product at the multiple predetermined time points Chromatogram.
  • the control parameters of each part of the distillation system are often set based on predetermined values, and it is impossible to dynamically adjust and optimize based on the actual situation.
  • there is a correlation between the control parameters of each part of the distillation system and the global optimization cannot be achieved by considering the conditions of each part alone, that is, the purity of the difluoromethane finally obtained cannot meet the preset requirements.
  • a gas chromatograph is used to obtain the gas chromatograms of the distillation products discharged from the distillation tower at the plurality of predetermined time points, and the gas chromatograms of the distillation products discharged from the distillation tower at the plurality of predetermined time points are obtained, and the gas chromatograms are Temperature and pressure sensors acquire multiple control parameters of the distillation system at multiple predetermined time points.
  • the multiple control parameters include: the pressure of the reflux tower, the temperature of the reflux tower, the degassing The pressure of the tower, the temperature of the degassing tower, the pressure of the rectification tower and the temperature of the rectification tower.
  • step S130 the gas chromatograms of the distillation products at the plurality of predetermined time points are passed through a first convolutional neural network using a three-dimensional convolution kernel to obtain a first feature map. That is, in the technical solution of the present application, the gas chromatograms of the distillation products at multiple predetermined time points are further processed through the first convolutional neural network using a three-dimensional convolution kernel to extract the The local correlation features of the gas chromatogram of the distillation product in the time series dimension are used to obtain the first feature map.
  • the first convolutional neural network using a three-dimensional convolution kernel can effectively extract the dynamic change characteristics of the distillation product.
  • steps S140 and S150 multiple control parameters at each predetermined time point are passed through a context encoder including an embedding layer to obtain multiple feature vectors, and the multiple feature vectors are concatenated to obtain Corresponding to the first feature vectors at each predetermined time point, the first feature vectors at each predetermined time point are then two-dimensionally arranged into a feature matrix and then passed through the second convolutional neural network to obtain the second feature map. It should be understood that there is a correlation between the control parameters of each part in the distillation system, and global optimization cannot be achieved by solely considering the conditions of each part.
  • multiple control parameters at each of the predetermined time points are further globally encoded through a context encoder including an embedding layer, so as to extract the high-dimensional implicit features of each parameter and each item.
  • a context encoder including an embedding layer
  • global high-dimensional latent features between parameters thereby obtaining multiple feature vectors.
  • multiple feature vectors can be concatenated to obtain the first feature vector corresponding to each predetermined time point, thereby facilitating subsequent feature extraction.
  • the first feature vectors at each predetermined time point are two-dimensionally arranged into a feature matrix and then processed through a second convolutional neural network to extract the hidden characteristics between the parameters between each time point.
  • sexually related features to obtain the second feature map.
  • the input data is subjected to convolution processing, pooling processing along the channel dimension and activation processing in the forward pass of the layer through each layer of the second convolutional neural network to be processed by the The last layer of the second convolutional neural network generates the second feature map, wherein the input of the first layer of the second convolutional neural network is the feature matrix.
  • step S160 the responsiveness estimate between the first feature map and the second feature map is calculated using normalization based on the characterization information relationship between the part and the whole to obtain a responsive feature map
  • the normalization based on the characterization information relationship between the local part and the whole is divided by the logarithmic function value of the sum of the feature value of each position in the first feature map and one by the second feature map
  • the logarithmic function value of the sum of the eigenvalues at all locations in and the sum of one can be further fused and passed through a classifier to obtain the desired control parameters. result.
  • the gas chromatography feature expressed by the first feature map F 1 is a local three-dimensional correlation of the three-dimensional convolution kernel Based on feature extraction, it focuses more on local feature expression, so it easily leads to low dependence on global responsiveness when calculating responsiveness. Therefore, in the technical solution of the present application, a normalized expression based on the characterization information relationship between the part and the whole is further used to calculate the responsiveness estimate between the first feature map and the second feature map to obtain the responsiveness.
  • Feature map since the gas chromatography feature expressed by the first feature map F 1 is a local three-dimensional correlation of the three-dimensional convolution kernel Based on feature extraction, it focuses more on local feature expression, so it easily leads to low dependence on global responsiveness when calculating responsiveness. Therefore, in the technical solution of the present application, a normalized expression based on the characterization information relationship between the part and the whole is further used to calculate the responsiveness estimate between the first feature map and the second feature map to obtain the responsiveness. Feature map.
  • the responsiveness feature map is passed through a classifier to obtain a classification result.
  • the classification result is used to indicate that the pressure of the rectification tower should be increased or decreased.
  • the temperature of the tower should be increased or should be decreased. That is to say, in the technical solution of the present application, after obtaining the response characteristic map, the response characteristic map is further passed through a classifier to obtain a value indicating whether the pressure of the distillation tower should be increased or decreased. Small, the classification result of the distillation tower temperature should be increased or should be reduced.
  • the classifier processes the responsiveness feature map with the following formula to generate a classification result, where the formula is: softmax ⁇ (W n ,B n ):...:( W 1 ,B 1 )
  • the control method of the distillation control system for the preparation of electronic-grade difluoromethane based on the embodiments of the present application is clarified, which uses a first convolutional neural network using a three-dimensional convolution kernel from multiple predetermined time points. Extract the dynamic change characteristics of the distillation product from the gas chromatogram of the distillation product, and use the context encoder to extract the high-dimensional implicit features of various control parameters at the multiple predetermined time points and among the various parameters The global high-dimensional latent features between the two feature information are further used to fuse the two feature information using a normalized expression based on the representation information relationship between the local and the whole. In this way, the robustness of minimizing the loss around the representation information is introduced to the responsiveness estimation.

Landscapes

  • Chemical & Material Sciences (AREA)
  • Organic Chemistry (AREA)
  • Chemical Kinetics & Catalysis (AREA)
  • Organic Low-Molecular-Weight Compounds And Preparation Thereof (AREA)

Abstract

一种用于电子级二氟甲烷制备的精馏控制系统(200)及其控制方法，所述用于电子级二氟甲烷制备的精馏控制系统(200)包括反应器(210)、回流塔(220)、脱气塔(230)、精馏塔(240)和控制器(250)。所述控制器(250)内部署有基于人工智能的参数控制算法以基于所述精馏控制系统(200)的全局情况来动态地调整所述精馏塔(240)的精馏控制参数，通过这样的方式，从优化控制的角度来提高电子级二氟甲烷的提纯精度。

Description

用于电子级二氟甲烷制备的精馏控制系统及其控制方法 技术领域
本发明涉及智能制造的领域，且更为具体地，涉及一种用于电子级二氟甲烷制备的精馏控制系统及其控制方法。
背景技术
二氟甲烷(R32)是一种热力学性能良好的氟利昂替代物。R32的ODP为0，GWP值很低，R32与其他组份形成的共沸物、近共沸混合物(如R407C、R410A等)被看作是最具潜力的R22的替代品。目前R32生产广泛使用液相氟化法，反应混合气中除主反应生成的HCl和R32外，还有中间产物R31，其他副反应生成的R23，R22，R40，R21，R143a，R50，以及夹带的少量原料R30和HF。
电子级的二氟甲烷的纯度高达99.9999％，这对现有的二氟甲烷的制备工艺提出了更高的要求。现有的厂商主要通过两个技术路线来提高二氟甲烷的纯度，一个技术方向为改变二氟甲烷的制备原理(本质上是化学方向)，另一个技术方向是优化二氟甲烷的提纯工艺(本质上是物理方向)，但不管是化学方向还是物理方向，最终都需要提纯工艺的优化(尤其是精馏工艺)才能够制备出满足纯度要求的电子级二氟甲烷。
在现有的精馏工艺中，精馏系统中各个部分的控制参数往往基于预定值来设定，无法基于实际的情况来动态调整优化。同时，精馏系统中各个部分的控制参数之间存在关联，单一考虑各个部分的情况无法做到全局最优，即，无法使得最终获得的二氟甲烷的纯度满足预设要求。因此，为了能够确保二氟甲烷的纯度满足预设要求，期待一种用于电子级二氟甲烷制备的精馏控制系统。
近年来，目前，深度学习以及神经网络已经广泛应用于计算机视觉、自然语言处理、语音信号处理等领域。此外，深度学习以及神经网络在图像分类、物体检测、语义分割、文本翻译等领域，也展现出了接近甚至超越人类的水平。
深度学习以及神经网络的发展为精馏塔的参数动态控制提供了新的解决思路和方案。
发明内容
为了解决上述技术问题，提出了本申请。本申请的实施例提供了一种用于电子级二氟甲烷制备的精馏控制系统及其控制方法，其中，所述用于电子级二氟甲烷制备的精馏控制系统包括反应器、回流塔、脱气塔、精馏塔和控制器。所述控制器内部署有基于人工智能的参数控制算法以基于所述精馏控制系统的全局情况来动态地调整所述精馏塔的精馏控制参数，通过这样的方式，从优化控制的角度来提高电子级二氟甲烷的提纯精度。
根据本申请的一个方面，提供了一种用于电子级二氟甲烷制备的精馏控制系统，其包括：
反应器，用于接收二氯甲烷和氟化氢，其中，所述二氯甲烷和所述氟化氢在催化剂的催化作用下发生反应以生成包含二氟甲烷的第一生成混合气，所述催化剂被装填于所述反应器内；
回流塔，用于接收所述包含二氟甲烷的第一生成混合气并从所述包含二氟甲烷的生成混合气中分离出所述氟化氢、所述二氟甲烷和所述一氟一氯甲烷以得到第二生成混合气；
脱气塔，用于接收所述第二生成混合气并除去所述第二生成混合气中的三氟甲烷和甲烷以得到第三生成混合气；
精馏塔，用于接收所述第三生成混合气并对所述第三生成混合气进行精馏以得到精馏产物，所述精馏产物为纯度大于等于99.9999％的电子级二氟甲烷；以及
控制器，用于基于所述精馏控制系统的全局参数来动态地控制所述精馏塔的温度和压力，所述精馏控制系统的全局参数包括所述回流塔的压力、所述回流塔的温度、所述脱气塔的压力、所述脱气塔的温度、所述精馏塔的压力和所述精馏塔的温度。
在根据本申请的用于电子级二氟甲烷制备的精馏控制系统中，所述控制器，用于：
获取多个预定时间点的所述精馏系统的多个控制参数，所述多个控制参数包括：所述回流塔的压力、所述回流塔的温度、所述脱气塔的压力、所述脱气塔的温度、所述精馏塔的压力和所述精馏塔的温度；
获取所述多个预定时间点的所述精馏产物的气相色谱图；
将所述多个预定时间点的精馏产物的气相色谱图通过使用三维卷积核 的第一卷积神经网络以获得第一特征图；
将各个所述预定时间点的多个控制参数通过包含嵌入层的上下文编码器以获得多个特征向量，并将多个特征向量进行级联以获得对应于各个预定时间点的第一特征向量；
将所述各个预定时间点的第一特征向量进行二维排列为特征矩阵后通过第二卷积神经网络以获得第二特征图；
使用基于局部和整体之间的表征信息关系的归一化来计算所述第一特征图和第二特征图之间的响应性估计以获得响应性特征图，其中，所述使用基于局部和整体之间的表征信息关系的归一化为以所述第一特征图中各个位置的特征值与一之和的对数函数值除以所述第二特征图中所有位置的特征值的求和与一之和的对数函数值；以及
将所述响应性特征图通过分类器以获得分类结果，所述分类结果用于表示所述精馏塔的压力应增大或应减小，所述精馏塔的温度应增大或应减小。
在根据本申请的用于电子级二氟甲烷制备的精馏控制系统中，所述控制器，进一步用于：使用所述使用三维卷积核的第一卷积神经网络以如下公式对所述多个预定时间点的精馏产物的气相色谱图进行编码以获得所述第一特征图；
其中，所述公式为：
Figure PCTCN2022097770-appb-000001
其中，H j、W j和R j分别表示三维卷积核的长度、宽度和高度，m表示第(i-1)层特征图的个数，
Figure PCTCN2022097770-appb-000002
是与(i-1)层的第m个特征图相连的卷积核，b ij为偏置，f表示激活函数。
在根据本申请的用于电子级二氟甲烷制备的精馏控制系统中，所述控制器，包括：
嵌入转化单元，用于使用所述包含嵌入层的上下文的编码器模型的嵌入层分别将各个所述预定时间点的多个控制参数转化为输入向量以获得参数输入向量的序列；
上下文编码单元，用于使用所述包含嵌入层的上下文的编码器模型的转换器对所述参数输入向量的序列进行基于全局的上下文语义编码以获得所 述多个特征向量；以及
级联单元，用于将多个特征向量进行级联以获得对应于各个预定时间点的第一特征向量。
在根据本申请的用于电子级二氟甲烷制备的精馏控制系统中，所述控制器，进一步用于：
将所述各个预定时间点的第一特征向量进行二维排列以获得特征矩阵；
通过所述第二卷积神经网络的各层在层的正向传递中分别对输入数据进行卷积处理、沿通道维度的池化处理和激活处理以由所述第二卷积神经网络的最后一层生成所述第二特征图，其中，所述第二卷积神经网络的第一层的输入为所述特征矩阵。
在根据本申请的用于电子级二氟甲烷制备的精馏控制系统中，所述控制器，进一步用于：
使用基于局部和整体之间的表征信息关系的归一化以如下公式来计算所述第一特征图和第二特征图之间的响应性估计以获得所述响应性特征图；
其中，所述公式为：
Figure PCTCN2022097770-appb-000003
其中
Figure PCTCN2022097770-appb-000004
和
Figure PCTCN2022097770-appb-000005
分别是所述第一特征图F 1、所述第二特征图F 2和所述响应性特征图的每个位置的特征值。
在根据本申请的用于电子级二氟甲烷制备的精馏控制系统中，所述控制器，进一步用于：使用所述分类器以如下公式对所述响应性特征图进行处理以生成分类结果；
其中，所述公式为：softmax{(W n,B n):…:(W 1,B 1)|Project(F)}，其中Project(F)表示将所述响应性特征图投影为向量，W 1至W n为各层全连接层的权重矩阵，B 1至B n表示各层全连接层的偏置矩阵。
根据本申请的另一方面，还提供了一种控制方法，其包括：
获取多个预定时间点的精馏系统的多个控制参数，所述多个控制参数包括：回流塔的压力、所述回流塔的温度、脱气塔的压力、所述脱气塔的温度、精馏塔的压力和所述精馏塔的温度；
获取所述多个预定时间点的精馏产物的气相色谱图；
将所述多个预定时间点的精馏产物的气相色谱图通过使用三维卷积核 的第一卷积神经网络以获得第一特征图；
将各个所述预定时间点的多个控制参数通过包含嵌入层的上下文编码器以获得多个特征向量，并将多个特征向量进行级联以获得对应于各个预定时间点的第一特征向量；
将所述各个预定时间点的第一特征向量进行二维排列为特征矩阵后通过第二卷积神经网络以获得第二特征图；
使用基于局部和整体之间的表征信息关系的归一化来计算所述第一特征图和第二特征图之间的响应性估计以获得响应性特征图，其中，所述使用基于局部和整体之间的表征信息关系的归一化为以所述第一特征图中各个位置的特征值与一之和的对数函数值除以所述第二特征图中所有位置的特征值的求和与一之和的对数函数值；以及
将所述响应性特征图通过分类器以获得分类结果，所述分类结果用于表示所述精馏塔的压力应增大或应减小，所述精馏塔的温度应增大或应减小。
在根据本申请的用于电子级二氟甲烷制备的精馏控制系统的控制方法中，将各个所述预定时间点的多个控制参数通过包含嵌入层的上下文编码器以获得多个特征向量，并将多个特征向量进行级联以获得对应于各个预定时间点的第一特征向量，包括：
使用所述包含嵌入层的上下文的编码器模型的嵌入层分别将各个所述预定时间点的多个控制参数转化为输入向量以获得参数输入向量的序列；
使用所述包含嵌入层的上下文的编码器模型的转换器对所述参数输入向量的序列进行基于全局的上下文语义编码以获得所述多个特征向量；以及
将多个特征向量进行级联以获得对应于各个预定时间点的第一特征向量。
在根据本申请的用于电子级二氟甲烷制备的精馏控制系统的控制方法中，使用基于局部和整体之间的表征信息关系的归一化来计算所述第一特征图和第二特征图之间的响应性估计以获得响应性特征图，包括：
使用基于局部和整体之间的表征信息关系的归一化以如下公式来计算所述第一特征图和第二特征图之间的响应性估计以获得响应性特征图；
其中，所述公式为：
Figure PCTCN2022097770-appb-000006
其中
Figure PCTCN2022097770-appb-000007
和
Figure PCTCN2022097770-appb-000008
分别是所述第一特征图F 1、所述第二特征图F 2和所述响应性特征图的每个位置的特征值。
与现有技术相比，本申请提供的用于电子级二氟甲烷制备的精馏控制系统及其控制方法，所述用于电子级二氟甲烷制备的精馏控制系统包括反应器、回流塔、脱气塔、精馏塔和控制器，其中，所述控制器内部署有基于人工智能的参数控制算法以基于所述精馏控制系统的全局情况来动态地调整所述精馏塔的精馏控制参数，通过这样的方式，从优化控制的角度来提高电子级二氟甲烷的提纯精度。
附图说明
通过结合附图对本申请实施例进行更详细的描述，本申请的上述以及其他目的、特征和优势将变得更加明显。附图用来提供对本申请实施例的进一步理解，并且构成说明书的一部分，与本申请实施例一起用于解释本申请，并不构成对本申请的限制。在附图中，相同的参考标号通常代表相同部件或步骤。
图1为根据本申请实施例的用于电子级二氟甲烷制备的精馏控制系统的框图示意图。
图2图示了根据本申请实施例的所述用于电子级二氟甲烷制备的精馏控制系统中控制器的框图示意图。
图3为根据本申请实施例的用于电子级二氟甲烷制备的精馏控制系统的控制方法的流程图。
图4为根据本申请实施例的用于电子级二氟甲烷制备的精馏控制系统的控制方法的架构示意图。
具体实施方式
下面，将参考附图详细地描述根据本申请的示例实施例。显然，所描述的实施例仅仅是本申请的一部分实施例，而不是本申请的全部实施例，应理解，本申请不受这里描述的示例实施例的限制。
场景概述
如前所述，二氟甲烷(R32)是一种热力学性能良好的氟利昂替代物。R32的ODP为0，GWP值很低，R32与其他组份形成的共沸物、近共沸混合物(如R407C、R410A等)被看作是最具潜力的R22的替代品。目前R32生产广泛使用液相氟化法，反应混合气中除主反应生成的HCl和R32外， 还有中间产物R31，其他副反应生成的R23，R22，R40，R21，R143a，R50，以及夹带的少量原料R30和HF。
电子级的二氟甲烷的纯度高达99.9999％，这对现有的二氟甲烷的制备工艺提出了更高的要求。现有的厂商主要通过两个技术路线来提高二氟甲烷的纯度，一个技术方向为改变二氟甲烷的制备原理(本质上是化学方向)，另一个技术方向是优化二氟甲烷的提纯工艺(本质上是物理方向)，但不管是化学方向还是物理方向，最终都需要提纯工艺的优化(尤其是精馏工艺)才能够制备出满足纯度要求的电子级二氟甲烷。
在现有的精馏工艺中，精馏系统中各个部分的控制参数往往基于预定值来设定，无法基于实际的情况来动态调整优化。同时，精馏系统中各个部分的控制参数之间存在关联，单一考虑各个部分的情况无法做到全局最优，即，无法使得最终获得的二氟甲烷的纯度满足预设要求。因此，为了能够确保二氟甲烷的纯度满足预设要求，期待一种用于电子级二氟甲烷制备的精馏控制系统。
近年来，目前，深度学习以及神经网络已经广泛应用于计算机视觉、自然语言处理、语音信号处理等领域。此外，深度学习以及神经网络在图像分类、物体检测、语义分割、文本翻译等领域，也展现出了接近甚至超越人类的水平。
深度学习以及神经网络的发展为精馏塔的参数动态控制提供了新的解决思路和方案。
具体地，在本申请的技术方案中，首先获取多个预定时间点的所述精馏系统的多个控制参数，所述多个控制参数包括：所述回流塔的压力、所述回流塔的温度、所述脱气塔的压力、所述脱气塔的温度、所述精馏塔的压力和所述精馏塔的温度，同时，获取所述多个预定时间点的所述精馏产物的气相色谱图。然后，将所述多个预定时间点的精馏产物的气相色谱图通过使用三维卷积核的第一卷积神经网络以获得第一特征图，特别地，使用三维卷积核的所述第一卷积神经网络能够有效地提取所述精馏产物的动态变化特征。同时，将各个所述预定时间点的多个控制参数通过包含嵌入层的上下文编码器以获得多个特征向量，并将多个特征向量进行级联以获得对应于各个预定时间点的第一特征向量。特别地，所述上下文编码器能够对各项参数进行基于上下文语义的全局编码以提取出各项参数的高维隐含特征和各项参数之间 的全局高维隐含特征。然后，将所述各个预定时间点的第一特征向量进行二维排列为特征矩阵后通过第二卷积神经网络以获得第二特征图，也就是，使用第二卷积神经网络来提取出各个时间点之间的参数之间的隐性关联。接着，融合第一特征图和第二特征图并通过分类器就可以获得当期待控制参数的控制结果。
然而，在计算所述第一特征图F 1对所述第二特征图F 2的响应性估计时，由于第一特征图F 1所表达的气相色谱特征是以三维卷积核的局部三维关联特征提取为基础的，其更聚焦于局部特征表达，因此在计算响应性时容易导致对全局响应性的依赖度低。基于此，使用基于局部和整体之间的表征信息关系的归一化表达来计算，具体为：
Figure PCTCN2022097770-appb-000009
其中
Figure PCTCN2022097770-appb-000010
和
Figure PCTCN2022097770-appb-000011
分别是第一特征图F 1、第二特征图F 2和响应性特征图的每个位置的特征值。
这样，通过向响应性估计引入围绕表征信息最小化损失的鲁棒性，来实现特征局部相当于特征整体的响应性的聚合性，从而提高响应性特征图对于所述第一特征图F 1对所述第二特征图F 2的期望响应性的全局依赖度，进而提高最终分类的准确度。
基于此，本申请提出了一种用于电子级二氟甲烷制备的精馏控制系统，其包括：反应器，用于接收二氟甲烷和氟化氢，其中，所述二氟甲烷和所述氟化氢在催化剂的催化作用下发生反应以生成包含二氟甲烷的第一生成混合气，所述催化剂被装填于所述反应器内；回流塔，用于接收所述包含二氟甲烷的第一生成混合气并从所述包含二氟甲烷的生成混合气中分离出所述氟化氢、所述二氟甲烷和所述一氟一氯甲烷以得到第二生成混合气；脱气塔，用于接收所述第二生成混合气并除去所述第二生成混合气中的三氟甲烷和甲烷以得到第三生成混合气；精馏塔，用于接收所述第三生成混合气并对所述第三生成混合气进行精馏以得到精馏产物，所述精馏产物为纯度大于等于99.9999％的电子级二氟甲烷；以及，控制器，用于：获取多个预定时间点的所述精馏系统的多个控制参数，所述多个控制参数包括：所述回流塔的压力、所述回流塔的温度、所述脱气塔的压力、所述脱气塔的温度、所述精馏塔的压力和所述精馏塔的温度；获取所述多个预定时间点的所述精馏产物的气相 色谱图；将所述多个预定时间点的精馏产物的气相色谱图通过使用三维卷积核的第一卷积神经网络以获得第一特征图；将各个所述预定时间点的多个控制参数通过包含嵌入层的上下文编码器以获得多个特征向量，并将多个特征向量进行级联以获得对应于各个预定时间点的第一特征向量；将所述各个预定时间点的第一特征向量进行二维排列为特征矩阵后通过第二卷积神经网络以获得第二特征图；使用基于局部和整体之间的表征信息关系的归一化来计算所述第一特征图和第二特征图之间的响应性估计以获得响应性特征图，其中，所述使用基于局部和整体之间的表征信息关系的归一化为以所述第一特征图中各个位置的特征值与一之和的对数函数值除以所述第二特征图中所有位置的特征值的求和与一之和的对数函数值；以及，将所述响应性特征图通过分类器以获得分类结果，所述分类结果用于表示所述精馏塔的压力应增大或应减小，所述精馏塔的温度应增大或应减小。
在介绍了本申请的基本原理之后，下面将参考附图来具体介绍本申请的各种非限制性实施例。
示例性系统
图2图示了根据本申请实施例的用于电子级二氟甲烷制备的精馏控制系统的流程图。如图2所示，根据本申请实施例的用于电子级二氟甲烷制备的精馏控制系统200，包括：反应器210、回流塔220、脱气塔230和精馏塔240和控制器250。
相应地，所述反应器210，用于接收二氯甲烷和氟化氢，其中，所述二氯甲烷和所述氟化氢在催化剂的催化作用下发生反应以生成包含二氟甲烷的第一生成混合气，所述催化剂被装填于所述反应器内。也就是说，所述反应器210是通过化学反应生成二氟甲烷的粗制品的场所。特别地，在本申请一个具体的示例张艳红，所述二氟甲烷采用气相氟化工艺制程，其化学反应过程包括：
CH 2Cl 2+HF→CH 2ClF+HCl
CH 2ClF+HF→CH 2F 2+HCL；
当然，在本申请其他示例中，所述二氟甲烷还可以通过其他原理制程，对此，并不为本申请所局限。
所述回流塔220，用于接收所述包含二氟甲烷的第一生成混合气并从所述包含二氟甲烷的生成混合气中分离出所述氟化氢、所述二氟甲烷和所述一 氟一氯甲烷以得到第二生成混合气。也就是说，从所述反应器210输出的所述所述包含二氟甲烷的第一生成混合气在输入所述回流塔220后，所述回流塔220通过其酸气分离系统依次分离出HF和HCl。
所述脱气塔230，用于接收所述第二生成混合气并除去所述第二生成混合气中的三氟甲烷和甲烷以得到第三生成混合气。具体地，考虑到所述三氟甲烷和甲烷具有相对较低的沸点，因此，在所述第二生成混合气进入所述脱气塔230内后，所述脱气塔能够通过利用所述第二生成混合气中不同成分具有不同的沸点来滤除低沸杂质三氟甲烷和甲烷(即，R23和R50)。
所述精馏塔240，用于接收所述第三生成混合气并对所述第三生成混合气进行精馏以得到精馏产物，所述精馏产物为纯度大于等于99.9999％的电子级二氟甲烷。也就是说，利用所述精馏塔240对所述第三生成混合气进行提纯以制得所述电子级二氟甲烷。
值得一提的是，在本申请实施例中，所述反应器210、所述回流塔220、所述脱气塔230和所述精馏塔240可采用任何现有的设备来构建所述精馏控制系统。相较于传统的精馏控制系统，本申请发明人从控制端的角度来优化一氟甲烷的精馏提纯精度。
如前所述，电子级的二氟甲烷的纯度高达99.9999％，这对现有的二氟甲烷的制备工艺提出了更高的要求。现有的厂商主要通过两个技术路线来提高二氟甲烷的纯度，一个技术方向为改变二氟甲烷的制备原理(本质上是化学方向)，另一个技术方向是优化二氟甲烷的提纯工艺(本质上是物理方向)，但不管是化学方向还是物理方向，最终都需要提纯工艺的优化(尤其是精馏工艺)才能够制备出满足纯度要求的电子级二氟甲烷。
然而，在现有的精馏工艺中，精馏系统中各个部分的控制参数往往基于预定值来设定，无法基于实际的情况来动态调整优化。同时，精馏系统中各个部分的控制参数之间存在关联，单一考虑各个部分的情况无法做到全局最优，即，无法使得最终获得的二氟甲烷的纯度满足预设要求。因此，为了能够确保二氟甲烷的纯度满足预设要求，本申请发明人尝试从优化控制的角度来提高电子级二氟甲烷的提纯精度。
具体地，如图1所示，根据本申请实施例的所述用于电子级二氟甲烷制备的精馏控制系统200，进一步包括所述控制器250，其中，所述控制器内部署有基于人工智能的参数控制算法以基于所述精馏控制系统的全局情况 来动态地调整所述精馏塔的精馏控制参数，通过这样的方式，从优化控制的角度来提高电子级二氟甲烷的提纯精度。
相应地，在本申请实施例中，所述控制器250，用于：获取多个预定时间点的所述精馏系统的多个控制参数，所述多个控制参数包括：所述回流塔的压力、所述回流塔的温度、所述脱气塔的压力、所述脱气塔的温度、所述精馏塔的压力和所述精馏塔的温度；获取所述多个预定时间点的所述精馏产物的气相色谱图；将所述多个预定时间点的精馏产物的气相色谱图通过使用三维卷积核的第一卷积神经网络以获得第一特征图；将各个所述预定时间点的多个控制参数通过包含嵌入层的上下文编码器以获得多个特征向量，并将多个特征向量进行级联以获得对应于各个预定时间点的第一特征向量；将所述各个预定时间点的第一特征向量进行二维排列为特征矩阵后通过第二卷积神经网络以获得第二特征图；使用基于局部和整体之间的表征信息关系的归一化来计算所述第一特征图和第二特征图之间的响应性估计以获得响应性特征图，其中，所述使用基于局部和整体之间的表征信息关系的归一化为以所述第一特征图中各个位置的特征值与一之和的对数函数值除以所述第二特征图中所有位置的特征值的求和与一之和的对数函数值；以及，将所述响应性特征图通过分类器以获得分类结果，所述分类结果用于表示所述精馏塔的压力应增大或应减小，所述精馏塔的温度应增大或应减小。
图2图示了根据本申请实施例的所述用于电子级二氟甲烷制备的精馏控制系统中控制器的框图示意图。如图2所示，所述控制器250，包括：参数获取单元251，用于获取多个预定时间点的所述精馏系统的多个控制参数，所述多个控制参数包括：所述回流塔的压力、所述回流塔的温度、所述脱气塔的压力、所述脱气塔的温度、所述精馏塔的压力和所述精馏塔的温度；产物数据获取单元252，用于获取所述多个预定时间点的所述精馏产物的气相色谱图；卷积编码单元253，用于将所述多个预定时间点的精馏产物的气相色谱图通过使用三维卷积核的第一卷积神经网络以获得第一特征图；上下文编码单元254，用于将各个所述预定时间点的多个控制参数通过包含嵌入层的上下文编码器以获得多个特征向量，并将多个特征向量进行级联以获得对应于各个预定时间点的第一特征向量；关联模式提取单元255，用于将所述各个预定时间点的第一特征向量进行二维排列为特征矩阵后通过第二卷积神经网络以获得第二特征图；多感受野归一化单元256，用于使用基于局 部和整体之间的表征信息关系的归一化来计算所述第一特征图和第二特征图之间的响应性估计以获得响应性特征图，其中，所述使用基于局部和整体之间的表征信息关系的归一化为以所述第一特征图中各个位置的特征值与一之和的对数函数值除以所述第二特征图中所有位置的特征值的求和与一之和的对数函数值；以及，控制结果生成单元257，用于将所述响应性特征图通过分类器以获得分类结果，所述分类结果用于表示所述精馏塔的压力应增大或应减小，所述精馏塔的温度应增大或应减小。
具体地，在本申请实施例中，在所述控制器250的参数获取单元251和所述产物数据获取单元252中，所述精馏控制系统的多个控制参数和精馏产物数据被采集。在具体实施中，可通过设置于精馏系统的多个温度和压力传感器获取多个预定时间点的所述精馏系统的多个控制参数，且通过气相色谱仪获取所述多个预定时间点的所述精馏塔排出的精馏产物的气相色谱图，也就是说，获取所述精馏控制系统的全局参数。
具体地，在本申请实施例中，在所述控制器250的卷积编码单元253，中，将所述多个预定时间点的精馏产物的气相色谱图通过使用三维卷积核的第一卷积神经网络以获得第一特征图。也就是，在本申请的技术方案中，进一步将所述多个预定时间点的精馏产物的气相色谱图通过使用三维卷积核的第一卷积神经网络中进行处理，以提取出所述精馏产物的气相色谱图在时序维度上的局部关联特征，从而获得第一特征图。特别地，这里，使用三维卷积核的所述第一卷积神经网络能够有效地提取所述精馏产物的动态变化特征。
更具体地，在本申请实施例中，将所述多个预定时间点的精馏产物的气相色谱图通过使用三维卷积核的第一卷积神经网络以获得第一特征图的过程，包括：以如下公式将所述多个预定时间点的精馏产物的气相色谱图通过使用三维卷积核的第一卷积神经网络以获得第一特征图；
其中，所述公式为：
Figure PCTCN2022097770-appb-000012
其中，H j、W j和R j分别表示三维卷积核的长度、宽度和高度，m表示第(i-1)层特征图的个数，
Figure PCTCN2022097770-appb-000013
是与(i-1)层的第m个特征图相连的卷积核， b ij为偏置，f表示激活函数。
具体地，在本申请实施例中，在所述控制器250的上下文编码单元254中，用于将各个所述预定时间点的多个控制参数通过包含嵌入层的上下文编码器以获得多个特征向量，并将多个特征向量进行级联以获得对应于各个预定时间点的第一特征向量。应可以理解，所述精馏系统中各个部分的控制参数之间存在关联，单一考虑各个所述部分的情况无法做到全局最优。因此，在本申请的技术方案中，进一步将各个所述预定时间点的多个控制参数通过包含嵌入层的上下文编码器中进行全局编码处理，以提取出各项参数的高维隐含特征和各项参数之间的全局高维隐含特征，从而获得多个特征向量。这样，就可以将多个特征向量进行级联以获得对应于各个预定时间点的第一特征向量，进而便于后续的特征提取。
更具体地，在本申请实施例中，将各个所述预定时间点的多个控制参数通过包含嵌入层的上下文编码器以获得多个特征向量，并将多个特征向量进行级联以获得对应于各个预定时间点的第一特征向量的过程，包括：首先，使用所述包含嵌入层的上下文的编码器模型的嵌入层分别将各个所述预定时间点的多个控制参数转化为输入向量以获得参数输入向量的序列。然后，使用所述包含嵌入层的上下文的编码器模型的转换器对所述参数输入向量的序列进行基于全局的上下文语义编码以获得所述多个特征向量。最后，将多个特征向量进行级联以获得对应于各个预定时间点的第一特征向量。
具体地，在本申请实施例中，在所述控制器250的关联模式提取单元255中，用于将所述各个预定时间点的第一特征向量进行二维排列为特征矩阵后通过第二卷积神经网络以获得第二特征图。也就是，在本申请的技术方案中，在得到所述第一特征向量后，将所述各个预定时间点的第一特征向量进行二维排列为特征矩阵后通过第二卷积神经网络中进行处理，以提取出所述各个时间点之间的参数之间的隐性关联特征，从而获得第二特征图。相应地，在一个具体示例中，通过所述第二卷积神经网络的各层在层的正向传递中对输入数据进行卷积处理、沿通道维度的池化处理和激活处理以由所述第二卷积神经网络的最后一层生成所述第二特征图，其中，所述第二卷积神经网络的第一层的输入为所述特征矩阵。
具体地，在本申请实施例中，在所述控制器250的多感受野归一化单元256中，用于使用基于局部和整体之间的表征信息关系的归一化来计算所述 第一特征图和第二特征图之间的响应性估计以获得响应性特征图，其中，所述使用基于局部和整体之间的表征信息关系的归一化为以所述第一特征图中各个位置的特征值与一之和的对数函数值除以所述第二特征图中所有位置的特征值的求和与一之和的对数函数值。应可以理解，在得到所述第一特征图和所述第二特征图后，就可以进一步融合所述第一特征图和所述第二特征图并通过分类器以获得当期待控制参数的控制结果。但是，在计算所述第一特征图F 1对所述第二特征图F 2的响应性估计时，由于第一特征图F 1所表达的气相色谱特征是以三维卷积核的局部三维关联特征提取为基础的，其更聚焦于局部特征表达，因此在计算响应性时容易导致对全局响应性的依赖度低。因此，在本申请的技术方案中，进一步使用基于局部和整体之间的表征信息关系的归一化表达来计算所述第一特征图和第二特征图之间的响应性估计以获得响应性特征图。
更具体地，在本申请实施例中，使用基于局部和整体之间的表征信息关系的归一化来计算所述第一特征图和第二特征图之间的响应性估计以获得响应性特征图的过程，包括：使用基于局部和整体之间的表征信息关系的归一化以如下公式来计算所述第一特征图和第二特征图之间的响应性估计以获得响应性特征图；
其中，所述公式为：
Figure PCTCN2022097770-appb-000014
其中
Figure PCTCN2022097770-appb-000015
和
Figure PCTCN2022097770-appb-000016
分别是所述第一特征图F 1、所述第二特征图F 2和所述响应性特征图的每个位置的特征值。应可以理解，这样，通过向所述响应性估计引入围绕表征信息最小化损失的鲁棒性，来实现特征局部相当于特征整体的响应性的聚合性，从而提高所述响应性特征图对于所述第一特征图F 1对所述第二特征图F 2的期望响应性的全局依赖度，进而提高分类的准确度。
具体地，在本申请实施例中，在所述控制器250的控制结果生成单元257中，用于将所述响应性特征图通过分类器以获得分类结果，所述分类结果用于表示所述精馏塔的压力应增大或应减小，所述精馏塔的温度应增大或应减小。也就是，在本申请的技术方案中，在得到所述响应性特征图后，进一步将所述响应性特征图通过分类器以获得用于表示所述精馏塔的压力应 增大或应减小，所述精馏塔的温度应增大或应减小的分类结果。相应地，在一个具体示例中，所述分类器以如下公式对所述响应性特征图进行处理以生成分类结果，其中，所述公式为：softmax{(W n,B n):…:(W 1,B 1)|Project(F)}，其中Project(F)表示将所述响应性特征图投影为向量，W 1至W n为各层全连接层的权重矩阵，B 1至B n表示各层全连接层的偏置矩阵。
综上，基于本申请实施例的所述用于电子级二氟甲烷制备的精馏控制系统200被阐明，其通过使用三维卷积核的第一卷积神经网络从多个预定时间点的所述精馏产物的气相色谱图中提取所述精馏产物的动态变化特征，并使用上下文编码器提取出所述多个预定时间点的各项控制参数的高维隐含特征和各项参数之间的全局高维隐含特征，进一步使用基于局部和整体之间的表征信息关系的归一化表达来融合这两个特征信息，这样，通过向响应性估计引入围绕表征信息最小化损失的鲁棒性，来实现特征局部相当于特征整体的响应性的聚合性，从而提高响应性特征图对于所述第一特征图对所述第二特征图的期望响应性的全局依赖度。进而，就能够提高分类的准确度。
如上所述，根据本申请实施例的用于电子级二氟甲烷制备的精馏控制系统200中的控制器250可以实现在各种终端设备中，例如用于电子级二氟甲烷制备的精馏控制算法的服务器等。在一个示例中，根据本申请实施例的所述控制器250可以作为一个软件模块和/或硬件模块而集成到终端设备中。例如，所述控制器250可以是该终端设备的操作系统中的一个软件模块，或者可以是针对于该终端设备所开发的一个应用程序；当然，所述控制器250同样可以是该终端设备的众多硬件模块之一。
替换地，在另一示例中，所述控制器250与该终端设备也可以是分立的设备，并且所述控制器250可以通过有线和/或无线网络连接到该终端设备，并且按照约定的数据格式来传输交互信息。
示例性方法
图3图示了用于电子级二氟甲烷制备的精馏控制系统的控制方法的流程图。如图3所示，根据本申请实施例的用于电子级二氟甲烷制备的精馏控制系统的控制方法，包括步骤：S110，获取多个预定时间点的所述精馏系统的多个控制参数，所述多个控制参数包括：所述回流塔的压力、所述回流塔的温度、所述脱气塔的压力、所述脱气塔的温度、所述精馏塔的压力和所述精馏塔的温度；S120，获取所述多个预定时间点的所述精馏产物的气相色谱图； S130，将所述多个预定时间点的精馏产物的气相色谱图通过使用三维卷积核的第一卷积神经网络以获得第一特征图；S140，将各个所述预定时间点的多个控制参数通过包含嵌入层的上下文编码器以获得多个特征向量，并将多个特征向量进行级联以获得对应于各个预定时间点的第一特征向量；S150，将所述各个预定时间点的第一特征向量进行二维排列为特征矩阵后通过第二卷积神经网络以获得第二特征图；S160，使用基于局部和整体之间的表征信息关系的归一化来计算所述第一特征图和第二特征图之间的响应性估计以获得响应性特征图，其中，所述使用基于局部和整体之间的表征信息关系的归一化为以所述第一特征图中各个位置的特征值与一之和的对数函数值除以所述第二特征图中所有位置的特征值的求和与一之和的对数函数值；以及，S170，将所述响应性特征图通过分类器以获得分类结果，所述分类结果用于表示所述精馏塔的压力应增大或应减小，所述精馏塔的温度应增大或应减小。
图4图示了根据本申请实施例的用于电子级二氟甲烷制备的精馏控制系统的控制方法的架构示意图。如图4所示，在所述用于电子级二氟甲烷制备的精馏控制系统的控制方法的网络架构中，首先，将获得的所述多个预定时间点的精馏产物的气相色谱图(例如，如图4中所示意的P1)通过使用三维卷积核的第一卷积神经网络(例如，如图4中所示意的CNN1)以获得第一特征图(例如，如图4中所示意的F1)；S140，将各个所述预定时间点的多个控制参数(例如，如图4中所示意的P2)通过包含嵌入层的上下文编码器(例如，如图4中所示意的E)以获得多个特征向量(例如，如图4中所示意的VF1)，并将多个特征向量进行级联以获得对应于各个预定时间点的第一特征向量(例如，如图4中所示意的VF2)；S150，将所述各个预定时间点的第一特征向量进行二维排列为特征矩阵(例如，如图4中所示意的MF)后通过第二卷积神经网络(例如，如图4中所示意的CNN2)以获得第二特征图(例如，如图4中所示意的F2)；S160，使用基于局部和整体之间的表征信息关系的归一化来计算所述第一特征图和第二特征图之间的响应性估计以获得响应性特征图(例如，如图4中所示意的F)；以及，最后，将所述响应性特征图通过分类器(例如，如图4中所示意的分类器)以获得分类结果，所述分类结果用于表示所述精馏塔的压力应增大或应减小，所述精馏塔的温度应增大或应减小。
更具体地，在步骤S110和S120中，获取多个预定时间点的所述精馏系 统的多个控制参数，所述多个控制参数包括：所述回流塔的压力、所述回流塔的温度、所述脱气塔的压力、所述脱气塔的温度、所述精馏塔的压力和所述精馏塔的温度，并获取所述多个预定时间点的所述精馏产物的气相色谱图。应可以理解，在现有的精馏工艺中，精馏系统中各个部分的控制参数往往基于预定值来设定，无法基于实际的情况来动态调整优化。同时，所述精馏系统中各个部分的控制参数之间存在关联，单一考虑各个部分的情况无法做到全局最优，即，无法使得最终获得的二氟甲烷的纯度满足预设要求。
因此，在本申请的技术方案中，首先，通过气相色谱仪获取所述多个预定时间点的所述精馏塔排出的精馏产物的气相色谱图，并且通过设置于精馏系统的多个温度和压力传感器获取多个预定时间点的所述精馏系统的多个控制参数，这里，所述多个控制参数包括：所述回流塔的压力、所述回流塔的温度、所述脱气塔的压力、所述脱气塔的温度、所述精馏塔的压力和所述精馏塔的温度。
更具体地，在步骤S130中，将所述多个预定时间点的精馏产物的气相色谱图通过使用三维卷积核的第一卷积神经网络以获得第一特征图。也就是，在本申请的技术方案中，进一步将所述多个预定时间点的精馏产物的气相色谱图通过使用三维卷积核的第一卷积神经网络中进行处理，以提取出所述精馏产物的气相色谱图在时序维度上的局部关联特征，从而获得第一特征图。特别地，这里，使用三维卷积核的所述第一卷积神经网络能够有效地提取所述精馏产物的动态变化特征。
更具体地，在步骤S140和步骤S150中，将各个所述预定时间点的多个控制参数通过包含嵌入层的上下文编码器以获得多个特征向量，并将多个特征向量进行级联以获得对应于各个预定时间点的第一特征向量，再将所述各个预定时间点的第一特征向量进行二维排列为特征矩阵后通过第二卷积神经网络以获得第二特征图。应可以理解，所述精馏系统中各个部分的控制参数之间存在关联，单一考虑各个所述部分的情况无法做到全局最优。因此，在本申请的技术方案中，进一步将各个所述预定时间点的多个控制参数通过包含嵌入层的上下文编码器中进行全局编码处理，以提取出各项参数的高维隐含特征和各项参数之间的全局高维隐含特征，从而获得多个特征向量。这样，就可以将多个特征向量进行级联以获得对应于各个预定时间点的第一特征向量，进而便于后续的特征提取。
进一步地，将所述各个预定时间点的第一特征向量进行二维排列为特征矩阵后通过第二卷积神经网络中进行处理，以提取出所述各个时间点之间的参数之间的隐性关联特征，从而获得第二特征图。相应地，在一个具体示例中，通过所述第二卷积神经网络的各层在层的正向传递中对输入数据进行卷积处理、沿通道维度的池化处理和激活处理以由所述第二卷积神经网络的最后一层生成所述第二特征图，其中，所述第二卷积神经网络的第一层的输入为所述特征矩阵。
更具体地，在步骤S160中，使用基于局部和整体之间的表征信息关系的归一化来计算所述第一特征图和第二特征图之间的响应性估计以获得响应性特征图，其中，所述使用基于局部和整体之间的表征信息关系的归一化为以所述第一特征图中各个位置的特征值与一之和的对数函数值除以所述第二特征图中所有位置的特征值的求和与一之和的对数函数值。应可以理解，在得到所述第一特征图和所述第二特征图后，就可以进一步融合所述第一特征图和所述第二特征图并通过分类器以获得当期待控制参数的控制结果。但是，在计算所述第一特征图F 1对所述第二特征图F 2的响应性估计时，由于第一特征图F 1所表达的气相色谱特征是以三维卷积核的局部三维关联特征提取为基础的，其更聚焦于局部特征表达，因此在计算响应性时容易导致对全局响应性的依赖度低。因此，在本申请的技术方案中，进一步使用基于局部和整体之间的表征信息关系的归一化表达来计算所述第一特征图和第二特征图之间的响应性估计以获得响应性特征图。应可以理解，这样，通过向所述响应性估计引入围绕表征信息最小化损失的鲁棒性，来实现特征局部相当于特征整体的响应性的聚合性，从而提高所述响应性特征图对于所述第一特征图F 1对所述第二特征图F 2的期望响应性的全局依赖度，进而提高分类的准确度。
更具体地，在步骤S170中，将所述响应性特征图通过分类器以获得分类结果，所述分类结果用于表示所述精馏塔的压力应增大或应减小，所述精馏塔的温度应增大或应减小。也就是，在本申请的技术方案中，在得到所述响应性特征图后，进一步将所述响应性特征图通过分类器以获得用于表示所述精馏塔的压力应增大或应减小，所述精馏塔的温度应增大或应减小的分类结果。相应地，在一个具体示例中，所述分类器以如下公式对所述响应性特征图进行处理以生成分类结果，其中，所述公式为： softmax{(W n,B n):…:(W 1,B 1)|Project(F)}，其中Project(F)表示将所述响应性特征图投影为向量，W 1至W n为各层全连接层的权重矩阵，B 1至B n表示各层全连接层的偏置矩阵。
综上，基于本申请实施例的所述用于电子级二氟甲烷制备的精馏控制系统的控制方法被阐明，其通过使用三维卷积核的第一卷积神经网络从多个预定时间点的所述精馏产物的气相色谱图中提取所述精馏产物的动态变化特征，并使用上下文编码器提取出所述多个预定时间点的各项控制参数的高维隐含特征和各项参数之间的全局高维隐含特征，进一步使用基于局部和整体之间的表征信息关系的归一化表达来融合这两个特征信息，这样，通过向响应性估计引入围绕表征信息最小化损失的鲁棒性，来实现特征局部相当于特征整体的响应性的聚合性，从而提高响应性特征图对于所述第一特征图对所述第二特征图的期望响应性的全局依赖度。进而，就能够提高分类的准确度。
以上结合具体实施例描述了本申请的基本原理，但是，需要指出的是，在本申请中提及的优点、优势、效果等仅是示例而非限制，不能认为这些优点、优势、效果等是本申请的各个实施例必须具备的。另外，上述公开的具体细节仅是为了示例的作用和便于理解的作用，而非限制，上述细节并不限制本申请为必须采用上述具体的细节来实现。

Claims (10)

  1. 一种用于电子级二氟甲烷制备的精馏控制系统，其特征在于，包括：反应器，用于接收二氯甲烷和氟化氢，其中，所述二氯甲烷和所述氟化氢在催化剂的催化作用下发生反应以生成包含二氟甲烷的第一生成混合气，所述催化剂被装填于所述反应器内；回流塔，用于接收所述包含二氟甲烷的第一生成混合气并从所述包含二氟甲烷的生成混合气中分离出所述氟化氢、所述二氟甲烷和所述一氟一氯甲烷以得到第二生成混合气；脱气塔，用于接收所述第二生成混合气并除去所述第二生成混合气中的三氟甲烷和甲烷以得到第三生成混合气；精馏塔，用于接收所述第三生成混合气并对所述第三生成混合气进行精馏以得到精馏产物，所述精馏产物为纯度大于等于99.9999％的电子级二氟甲烷；以及控制器，用于基于所述精馏控制系统的全局参数来动态地控制所述精馏塔的温度和压力，所述精馏控制系统的全局参数包括所述回流塔的压力、所述回流塔的温度、所述脱气塔的压力、所述脱气塔的温度、所述精馏塔的压力和所述精馏塔的温度。
  2. 根据权利要求1所述的用于电子级二氟甲烷制备的精馏控制系统，其中，所述控制器，用于：获取多个预定时间点的所述精馏系统的多个控制参数，所述多个控制参数包括：所述回流塔的压力、所述回流塔的温度、所述脱气塔的压力、所述脱气塔的温度、所述精馏塔的压力和所述精馏塔的温度；获取所述多个预定时间点的所述精馏产物的气相色谱图；将所述多个预定时间点的精馏产物的气相色谱图通过使用三维卷积核的第一卷积神经网络以获得第一特征图；将各个所述预定时间点的多个控制参数通过包含嵌入层的上下文编码器以获得多个特征向量，并将多个特征向量进行级联以获得对应于各个预定时间点的第一特征向量；将所述各个预定时间点的第一特征向量进行二维排列为特征矩阵后通过第二卷积神经网络以获得第二特征图；使用基于局部和整体之间的表征信息关系的归一化来计算所述第一特征图和第二特征图之间的响应性估计以获得响应性特征图，其中，所述使用基于局部和整体之间的表征信息关系的归一化为以所述第一特征图中各个位置的特征值与一之和的对数函数值除以所述第二特征图中所有位置的特征值的求和与一之和的对数函数值；以及将所述响应性特征图通过分类器以获得分类结果，所述分类结果用于表示所述精馏塔的压力应增大或应减小，所述精馏塔的温度应增大或应减小。
  3. 根据权利要求2所述的用于电子级二氟甲烷制备的精馏控制系统，其中，所述控制器，进一步用于：使用所述使用三维卷积核的第一卷积神经网络以如下公式对所述多个预定时间点的精馏产物的气相色谱图进行编码以获得所述第一特征图；其中，所述公式为：
    Figure PCTCN2022097770-appb-100001
    其中，H j、W j和R j分别表示三维卷积核的长度、宽度和高度，m表示第(i-1)层特征图的个数，
    Figure PCTCN2022097770-appb-100002
    是与(i-1)层的第m个特征图相连的卷积核，b ij为偏置，f表示激活函数。
  4. 根据权利要求2所述的用于电子级二氟甲烷制备的精馏控制系统，其中，所述控制器，包括：嵌入转化单元，用于使用所述包含嵌入层的上下文的编码器模型的嵌入层分别将各个所述预定时间点的多个控制参数转化为输入向量以获得参数输入向量的序列；上下文编码单元，用于使用所述包含嵌入层的上下文的编码器模型的转换器对所述参数输入向量的序列进行基于全局的上下文语义编码以获得所述多个特征向量；以及级联单元，用于将多个特征向量进行级联以获得对应于各个预定时间点的第一特征向量。
  5. 根据权利要求4所述的用于电子级二氟甲烷制备的精馏控制系统，其中，所述控制器，进一步用于：将所述各个预定时间点的第一特征向量进行二维排列以获得特征矩阵；通过所述第二卷积神经网络的各层在层的正向传递中分别对输入数据进行卷积处理、沿通道维度的池化处理和激活处理以由所述第二卷积神经网络的最后一层生成所述第二特征图，其中，所述第二卷积神经网络的第一层的输入为所述特征矩阵。
  6. 根据权利要求5所述的用于电子级二氟甲烷制备的精馏控制系统，其中，所述控制器，进一步用于：使用基于局部和整体之间的表征信息关系的归一化以如下公式来计算所述第一特征图和第二特征图之间的响应性估计以获得所述响应性特征图；其中，所述公式为：
    Figure PCTCN2022097770-appb-100003
    其中
    Figure PCTCN2022097770-appb-100004
    和
    Figure PCTCN2022097770-appb-100005
    分别是所述第一特征图F 1、所述第二特征图F 2和所述响应性特征图的每个位置的特征值。
  7. [根据细则91更正 30.06.2022]
    根据权利要求6所述的用于电子级二氟甲烷制备的精馏控制系统，其中，所述控制器，进一步用于：使用所述分类器以如下公式对所述响应性特征图进行处理以生成分类结果；其中，所述公式为： softmax{(W n,B n):…:(W 1,B 1)|Project(F)}，其中Project(F)表示将所述响应性特征图投影为向量，W 1至W n为各层全连接层的权重矩阵，B 1至B n表示各层全连接层的偏置矩阵。
  8. 一种控制方法，其特征在于，包括：获取多个预定时间点的精馏系统的多个控制参数，所述多个控制参数包括：回流塔的压力、所述回流塔的温度、脱气塔的压力、所述脱气塔的温度、精馏塔的压力和所述精馏塔的温度；获取所述多个预定时间点的精馏产物的气相色谱图；将所述多个预定时间点的精馏产物的气相色谱图通过使用三维卷积核的第一卷积神经网络以获得第一特征图；将各个所述预定时间点的多个控制参数通过包含嵌入层的上下文编码器以获得多个特征向量，并将多个特征向量进行级联以获得对应于各个预定时间点的第一特征向量；将所述各个预定时间点的第一特征向量进行二维排列为特征矩阵后通过第二卷积神经网络以获得第二特征图；使用基于局部和整体之间的表征信息关系的归一化来计算所述第一特征图和第二特征图之间的响应性估计以获得响应性特征图，其中，所述使用基于局部和整体之间的表征信息关系的归一化为以所述第一特征图中各个位置的特征值与一之和的对数函数值除以所述第二特征图中所有位置的特征值的求和与一之和的对数函数值；以及将所述响应性特征图通过分类器以获得分类结果，所述分类结果用于表示所述精馏塔的压力应增大或应减小，所述精馏塔的温度应增大或应减小。
  9. 根据权利要求8所述的用于电子级二氟甲烷制备的精馏控制系统的控制方法，其中，将各个所述预定时间点的多个控制参数通过包含嵌入层的上下文编码器以获得多个特征向量，并将多个特征向量进行级联以获得对应于各个预定时间点的第一特征向量，包括：使用所述包含嵌入层的上下文的编码器模型的嵌入层分别将各个所述预定时间点的多个控制参数转化为输入向量以获得参数输入向量的序列；使用所述包含嵌入层的上下文的编码器模型的转换器对所述参数输入向量的序列进行基于全局的上下文语义编码以获得所述多个特征向量；以及将多个特征向量进行级联以获得对应于各个预定时间点的第一特征向量。
  10. 根据权利要求8所述的用于电子级二氟甲烷制备的精馏控制系统的控制方法，其中，使用基于局部和整体之间的表征信息关系的归一化来计算所述第一特征图和第二特征图之间的响应性估计以获得响应性特征图，包括：
    使用基于局部和整体之间的表征信息关系的归一化以如下公式来计算所述第一特征图和第二特征图之间的响应性估计以获得响应性特征图；
    其中，所述公式为：
    Figure PCTCN2022097770-appb-100006
    其中
    Figure PCTCN2022097770-appb-100007
    和
    Figure PCTCN2022097770-appb-100008
    分别是所述第一特征图F 1、所述第二特征图F 2和所述响应性特征图的每个位置的特征值。
PCT/CN2022/097770 2022-04-29 2022-06-09 用于电子级二氟甲烷制备的精馏控制系统及其控制方法 WO2023206724A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210465413.1A CN114768279B (zh) 2022-04-29 2022-04-29 用于电子级二氟甲烷制备的精馏控制系统及其控制方法
CN202210465413.1 2022-04-29

Publications (1)

Publication Number Publication Date
WO2023206724A1 true WO2023206724A1 (zh) 2023-11-02

Family

ID=82435571

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/097770 WO2023206724A1 (zh) 2022-04-29 2022-06-09 用于电子级二氟甲烷制备的精馏控制系统及其控制方法

Country Status (2)

Country Link
CN (1) CN114768279B (zh)
WO (1) WO2023206724A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117180952A (zh) * 2023-11-07 2023-12-08 湖南正明环保股份有限公司 多向气流料层循环半干法烟气脱硫系统及其方法

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115202265A (zh) * 2022-07-29 2022-10-18 福建天甫电子材料有限公司 电子级氢氧化钾的智慧产线的控制系统及其控制方法
CN115238591B (zh) * 2022-08-12 2022-12-27 杭州国辰智企科技有限公司 动态参数校验与驱动cad自动建模引擎系统
CN115599049B (zh) * 2022-08-31 2023-04-07 福建省龙氟新材料有限公司 用于无水氟化氢生产的能源管理控制系统及其控制方法
CN115688592B (zh) * 2022-11-09 2023-05-09 福建德尔科技股份有限公司 用于电子级四氟化碳制备的精馏控制系统及其方法
CN116825217B (zh) * 2023-03-15 2024-05-14 福建省德旭新材料有限公司 制备高纯五氟化磷的方法

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1962015A (zh) * 2006-10-30 2007-05-16 浙江大学 高纯精馏的动态矩阵控制系统和方法
CN101073712A (zh) * 2006-12-26 2007-11-21 浙江大学 基于广义预测控制的精馏塔高纯度精馏控制系统及方法
CN102339040A (zh) * 2010-07-15 2012-02-01 清华大学 精馏塔优化控制方法
CN104635493A (zh) * 2015-01-13 2015-05-20 中国石油大学(华东) 基于温度波模型预测控制的内部热耦合精馏控制装置
CN107261541A (zh) * 2017-08-23 2017-10-20 广州百兴网络科技有限公司 一种精馏装置及精馏控制方法
CN108929193A (zh) * 2018-06-28 2018-12-04 江苏三美化工有限公司 一种高纯度二氟甲烷的精馏工艺
US20200108327A1 (en) * 2018-10-08 2020-04-09 Uop Llc High Purity Distillation Process Control
CN112919419A (zh) * 2021-01-29 2021-06-08 福建德尔科技有限公司 电子级三氟化氯的精馏纯化系统控制方法

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP4134391B2 (ja) * 1998-04-07 2008-08-20 日本ゼオン株式会社 不飽和炭化水素の分離精製装置および分離精製方法
CN111144490B (zh) * 2019-12-26 2022-09-06 南京邮电大学 一种基于轮替知识蒸馏策略的细粒度识别方法
KR102139358B1 (ko) * 2020-04-22 2020-07-29 한국생산기술연구원 머신러닝 기반 플랫폼을 이용한 공정제어방법, 그를 수행하기 위한 컴퓨터 프로그램 매체 및 공정제어장치
AU2020104006A4 (en) * 2020-12-10 2021-02-18 Naval Aviation University Radar target recognition method based on feature pyramid lightweight convolutional neural network
CN113987937A (zh) * 2021-10-27 2022-01-28 北京航空航天大学 基于卷积神经网络的热强化sve有害气体浓度检测方法

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1962015A (zh) * 2006-10-30 2007-05-16 浙江大学 高纯精馏的动态矩阵控制系统和方法
CN101073712A (zh) * 2006-12-26 2007-11-21 浙江大学 基于广义预测控制的精馏塔高纯度精馏控制系统及方法
CN102339040A (zh) * 2010-07-15 2012-02-01 清华大学 精馏塔优化控制方法
CN104635493A (zh) * 2015-01-13 2015-05-20 中国石油大学(华东) 基于温度波模型预测控制的内部热耦合精馏控制装置
CN107261541A (zh) * 2017-08-23 2017-10-20 广州百兴网络科技有限公司 一种精馏装置及精馏控制方法
CN108929193A (zh) * 2018-06-28 2018-12-04 江苏三美化工有限公司 一种高纯度二氟甲烷的精馏工艺
US20200108327A1 (en) * 2018-10-08 2020-04-09 Uop Llc High Purity Distillation Process Control
CN112919419A (zh) * 2021-01-29 2021-06-08 福建德尔科技有限公司 电子级三氟化氯的精馏纯化系统控制方法

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117180952A (zh) * 2023-11-07 2023-12-08 湖南正明环保股份有限公司 多向气流料层循环半干法烟气脱硫系统及其方法
CN117180952B (zh) * 2023-11-07 2024-02-02 湖南正明环保股份有限公司 多向气流料层循环半干法烟气脱硫系统及其方法

Also Published As

Publication number Publication date
CN114768279A (zh) 2022-07-22
CN114768279B (zh) 2022-11-11

Similar Documents

Publication Publication Date Title
WO2023206724A1 (zh) 用于电子级二氟甲烷制备的精馏控制系统及其控制方法
CN114870416B (zh) 用于电子级一氟甲烷制备的精馏控制系统及精馏控制方法
WO2024045244A1 (zh) 用于无水氟化氢生产的能源管理控制系统及其控制方法
WO2024000798A1 (zh) 用于电子级氢氟酸制备的生产管理控制系统及其控制方法
CN115688592B (zh) 用于电子级四氟化碳制备的精馏控制系统及其方法
CN115356434B (zh) 六氟丁二烯储放场所的气体监测系统及其监测方法
WO2023226228A1 (zh) 用于电子级双氧水制备的智能化有毒有害气体报警系统
WO2024000800A1 (zh) 用于六氟磷酸锂制备的能源管理控制系统及其控制方法
CN115231525B (zh) 电子级三氟化氯的智能分离纯化系统
CN115309215B (zh) 氟化铵制备用的自动配料控制系统及其控制方法
WO2023226236A1 (zh) 用于电子级氢氟酸制备的能源管理控制系统及其控制方法
CN115291646B (zh) 用于氟化锂制备的能源管理控制系统及其控制方法
CN116047987B (zh) 用于电子级缓冲氧化物蚀刻液生产的智能控制系统
CN115090200B (zh) 用于电子级氢氟酸制备的自动配料系统及其配料方法
CN115240046A (zh) 用于缓冲氧化物蚀刻液生产的自动配料系统及其配料方法
WO2024000828A1 (zh) 用于光阻剥离液生产的自动配料系统及其配料方法
WO2024113599A1 (zh) 用于电子级六氟丁二烯制备的生产管理控制系统
CN115845428A (zh) 用于诱导六氟磷酸结晶的超声波装置及其控制方法
CN115841644A (zh) 基于物联网的城市基础建设工程设备的控制系统及其方法
CN115601318A (zh) 快吸收低反渗纸尿裤智能生产方法及其系统
CN115754108B (zh) 一种电子级六氟丁二烯的酸度测定系统及其方法
WO2024036690A1 (zh) 用于剥膜液生产的自动配料系统及其配料方法
CN108929193A (zh) 一种高纯度二氟甲烷的精馏工艺
CN112960213A (zh) 使用特征概率分布表示的智能包装质量检测方法
CN117455451A (zh) 一种草甘膦水解溶剂的回收方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22939556

Country of ref document: EP

Kind code of ref document: A1